Planning for Task Assignment Agility in Large Heterogeneous Teams Session 09 Paper 087

1. James C. Kilian, Ph.D. Systems Engineer, Principal Lockheed Martin Maritime Systems & Sensors 77 S. Bedford Rd, Burlington, MA 01803 781-565-1059 james.c.kilian@lmco.com Planning for Task Assignment Agility in Large Heterogeneous TeamsSession 09Paper 087 06/18/08 LOCKHEED MARTIN PROPRIETARY INFORMATION

2. A Complex Endeavor • Problem Domain: Planning and coordination of a large heterogeneous team in an adversarial situation • The state of operating environment is dynamic and uncertain and only partially observable with finite horizon • The adversary’s tactics may have unpreditable elements and are only partially observable with finite horizon • Each asset on the team may not be 100% available or reliable • Asset deployment is assumed to be a singular event • Examples • Intrusion prevention • Missile defense • Blockades • Force protection • Air defense • Asymmetric conflict

3. Complexity & Agility • The distinction between complicated and complex endeavors has been made†. • The former is defined as a situation in which causal relations are well-understood and thus are directly available for analysis. • In a complex situation, changes to the situation are uncertain (though not purely random). They may also be characterized by instability. • Agility in planning and decision-making systems addresses complex situations • Properties of agile planning includes • Minimizing irretrievable commitment of resources • Improve information position • Shape the adversary’s information position • Engage in situated planning/decision making † Alberts, D.S. and R.E. Hayes, 2007, Planning: Complex Endeavors, (Washington D.C. CCRP).

4. Agility in Planning • General Approaches • Conditional planning (contingencies) • Monitor and Repair (re)planning • Continual planning • Distributed Multi-Agent Approaches • Cooperative distributed planning • Negotiated distributed planning A continual cooperative distributed planning approach is examined here

5. A 3-Stage Mission Planning Model Deliberative Processing Plan Initialization Threat Event Reported Threat Prosecuted Assets Deployed Plan Refinement no no no yes yes yes Reactive Processing Incremental Adaptation Direct asset allocation (First 2 stages) to achieve agility in task allocation (3rd stage)

6. Getting to Task Assignment Agility • Task assignment agility therefore depends on • Defining a sufficient team of assets for which there is at least one task assignment configuration for which the team meets its performance requirements • Minimizing single point failures in both planning and plan execution • Achieving suitable plan update response times • Separate the planning problem into asset distribution and task assignment configuration subproblems • Model-based distributed co-evolutionary computation to minimize single point planning failures and achieve computational efficiency

7. Optimizing Asset Distribution A static assignment matrix: assets (rows) tasks (columns) p(x,y) is a penalty function accommodating constraints in the problem† † Kuri, A. 1998, in Proc. 6th European Congress on Intelligent Techniques and Soft Computing. This objective function preserves the most flexibility for task assignment configurations

8. Iterative Factored Optimization I Conditional Optimization within a population of subspaces 2D Optimization Problem         For each given but arbitrary value of y find optimal x For each given but arbitrary value of x find optimal y At each step the optimization search is only over the subspaces defined by the complementary process, so there is no explicit global search These lines correspond to the results for the complementary variable from the conditional optimizations in the previous step       Update the condition variable for each subproblem and repeat. As can be seen x and y have found the peak in the graph

9. Evolutionary Agent Evolutionary Agent x1 x2 Evolutionary Agent Evolutionary Agent x3 x4 Iterative Factored Optimization II • An approach to concurrent subspace optimization is to co-evolve the states of loosely-coupled agents† • This technique has been applied to planning problems in the “agile” manufacturing arena • The technique is also similar to so-called “covering problems” Primary Search Variables Network Co-Evolution – The Modified Potter Architecture † Potter, M.A., 1997, The Design and Analysis of a Computational Model for Cooperative Co-evolution, Ph.D. Thesis (George Mason University).

10. An Architecture for Model Based Evolutionary Optimization Set Planning Conditions Build Probabilistic Model Initialize Risk distribution Update Probabilistic Model Sample Model to Create New Population No Converged? Save Data Plot Results Calculate Fitness Yes Model-Based Genetic Algorithms (MBGA) have o(nln(n)) complexity whereas standard GA algorithms have complexity ranging from o(nln(n)) to o(en)† † M. Pelikan, 2005, Hierarchical Bayesian Optimization Algorithm, (Springer).

11. Asset Asset Detect Track Shoot Support Detect Track Shoot Support Subscription Notification Allocation Capabilities Capabilities Capabilities SOA Context – Decentralized Coordination This architecture is based on a publish-subscribe pattern as opposed to the query-response pattern (Technically making this model an event-driven architecture (EDA) rather than a strict SOA) Task Service Layer Asset Entity Service Layer Detect Track Shoot Support Utility Service Layer Co-evolving ensemble configuration in a network service model

12. A Toy Example – Intrusion Detection • A circular region with discrete entry points • One or more intruders may traverse the region in a straight line connecting any two non-adjacent entry points • A set of three assets charged with detecting and intercepting intruders • Spatial constraints confine the assets to points on a grid and limit the extents to which an asset can achieve detection or interception 1 1 3 2 2 3

13. Summary • Demonstrated a realization of agility in response to a complex endeavor through the integration of three concepts: • An asset commitment objective function that preserves the maximum number of choices for subsequent task assignment configuration optimization • Nonhierarchical Distributed Optimization through co-evolution • Load balancing across multiple assets naturally accommodated • Scalable with increasing numbers of assets • Global optimality is an emergent characteristic (consistent with a system-of-systems POV) • Model-Based Optimization – Builds a model within an evolutionary (stochastic optimization) process • Scalability well-suited to managing complexity • Posterior distribution allows assessment of plan robustness • Situated Task Assignment Optimization – The ability to adapt a plan to unforeseen events as they unfold • better supports collaboration • better at handling uncertainty than classical planning